Decision-Making of automated vehicles under diverse risky pedestrian crossing behaviors

Uncontrolled midblocks are frequently associated with elevated traffic conflict rates but often lack effective mitigation measures. Risky pedestrian crossing behaviors, such as jaywalking, distracted walking, and dart-outs from occluded areas, combined with heterogeneous driving styles, further complicate automated vehicle (AV) decision-making. However, most existing studies focus on simplified scenarios and rarely consider complex settings. This gap limits the realism and applicability of current AV decision-making research in urban environments. To address these challenges, a high-fidelity multi-agent simulation platform replicates the dynamic interactions among AVs, human-driven vehicles, and pedestrians with diverse risky crossing behaviors. A general visibility modeling method using polar-sector analysis simulates perceptual limitations caused by occlusions for multiple agents. On this basis, a deep reinforcement learning (DRL)-based decision-making framework is developed to integrate risk assessment with safety filtering. The framework dynamically incorporates environmental risk into the behavior policy of AVs and, during execution, employs a safety filter to correct or replace unsafe actions. Experimental results show that the proposed approach substantially improves safety margins and control smoothness in complex scenarios with occluded or distracted pedestrians. Compared to rule-based and risk-unaware DRL baselines, the learned policy exhibits stronger anticipatory behavior and achieves a better balance between safety and traffic efficiency. These findings highlight the promise of risk-aware DRL for managing highly uncertain and interactive urban driving environments. The approach provides new insights for the safe deployment of AVs in real-world traffic.